-
Members: FreeCIS
IEEE Members: Free
Non-members: FreeLength: 00:55:07
08 Jul 2018
Mengjie Zhang Keynote talk (CEC) at WCCI 2018
Abstract: "Image analysis problems occur in our everyday life. Recognising faces in digital images and diagnosing medical conditions from X-Ray images are just two examples of the many important tasks for which we need computer based image analysis systems. Since the 1980s, many image analysis algorithms have been developed. Among those algorithms, deep learning particularly deep convolutional neural networks have received very good success and attracted attentions to industry people and researchers in computer vision and image processing, neural networks, and machine learning. However, there are at least three major limitations in deep convolutional neural networks: (1) the learning architecture including the number of layers, the number of feature maps in each layer and the number of nodes in each feature map are still very much determined manually via ""trial and error"", which requires a large amount of hand-crafting/trial time and good domain knowledge. However, such experts are hard to find in many cases, or using such expertise is too expensive. (2) Almost all the current deep learning algorithms need a large number of examples/instances (e.g. AlphaGo used over 30 million instances) that many problems do not have. (3) Those algorithms require a huge computational cost that big companies such as Google, Baidu, and Microsoft can cope well but most universities and research institutions cannot.
To address these limitations, evolutionary computation techniques start playing a significant role for automatically determining deep structures, transfer functions and parameters to tackle image analysis tasks, and have great potential to advance the developments of deep structures and algorithms. This talk will provide an extended view of deep learning, overview the state-of-the-art work in evolutionary deep learning using GAs/PSO/DE, and discuss some recent developments using Genetic Programming (GP) to automatically evolving deep structures and feature construction for image recognition with a highlight of the interpretation capability and visualisation of constructed features. Finally, recent work and ideas on evolutionary deep transfer learning will be discussed.
"
Abstract: "Image analysis problems occur in our everyday life. Recognising faces in digital images and diagnosing medical conditions from X-Ray images are just two examples of the many important tasks for which we need computer based image analysis systems. Since the 1980s, many image analysis algorithms have been developed. Among those algorithms, deep learning particularly deep convolutional neural networks have received very good success and attracted attentions to industry people and researchers in computer vision and image processing, neural networks, and machine learning. However, there are at least three major limitations in deep convolutional neural networks: (1) the learning architecture including the number of layers, the number of feature maps in each layer and the number of nodes in each feature map are still very much determined manually via ""trial and error"", which requires a large amount of hand-crafting/trial time and good domain knowledge. However, such experts are hard to find in many cases, or using such expertise is too expensive. (2) Almost all the current deep learning algorithms need a large number of examples/instances (e.g. AlphaGo used over 30 million instances) that many problems do not have. (3) Those algorithms require a huge computational cost that big companies such as Google, Baidu, and Microsoft can cope well but most universities and research institutions cannot.
To address these limitations, evolutionary computation techniques start playing a significant role for automatically determining deep structures, transfer functions and parameters to tackle image analysis tasks, and have great potential to advance the developments of deep structures and algorithms. This talk will provide an extended view of deep learning, overview the state-of-the-art work in evolutionary deep learning using GAs/PSO/DE, and discuss some recent developments using Genetic Programming (GP) to automatically evolving deep structures and feature construction for image recognition with a highlight of the interpretation capability and visualisation of constructed features. Finally, recent work and ideas on evolutionary deep transfer learning will be discussed.
"
